CI: contract + canary + live e2e suites (channels, voice)

[dimavrem22]

Ports the proven plugin-CI stack (hermes-agent-plugin, openclaw-plugin) to this bridge. Same three tiers, adapted to a Python bridge whose host is the Claude Code CLI + claude-agent-sdk:

Tiers

1. tests.yml — offline unit suite on every push/PR (Python 3.10/3.12 matrix; absorbs the old ci.yml), plus a contract-pr job that runs tests/contract against the latest published claude-agent-sdk and @anthropic-ai/claude-code so upstream interface drift fails the PR, not a live gateway.
2. canary.yml — the same contract suite 2×/day (06:13 / 18:13 PT — deliberately an hour ahead of the rest of the fleet's canaries), paging Google Chat only on unattended (scheduled) failure.
3. Live e2e — boots the real bridge (tunnel + webhooks + Claude Code sessions) in the runner and drives it from the remote driver identity:
   • live-channels.yml: email reachability, email intelligence (identity / sender / tools / contact CRUD), SMS, and cross-channel (email→SMS, SMS→email, email→call, SMS→call). Two serialized legs: mock (deterministic local Anthropic-API mock via ANTHROPIC_BASE_URL — free, proves the whole pipe with a nonce echo) and real (real Claude model — proves reasoning + tool use). Chained off a passing canary via workflow_run.
   • live-voice.yml: two real phone-call scenarios — inbound_inkbox (driver calls the agent; Inkbox STT/TTS) and outbound_realtime (driver texts "call me"; the agent dials back on the realtime voice path). Transcript-verified two-way speech + speech-mode assertions on the AUT's call records.

External-events live tests are intentionally out of scope, matching the fleet decision in inkbox-ai/openclaw-plugin#24.

Gating & safety

• Live suites run only on ready (non-draft) same-repo PRs, manual dispatch, or a passing canary; drafts and forks get the offline lanes only.
• Repo-wide inkbox-live-aut-tunnel concurrency group (no cancel) — only one holder of the AUT identity's Inkbox tunnel at a time, across both live workflows.
• Gateway/driver logs can carry live message content, so they're dumped and uploaded only on failure.
• INKBOX_PERMISSION_TIMEOUT_S=30 in CI so a stray permission escalation fails fast instead of parking a session for 10 minutes.

Secrets

Uses the existing CLAUDE_CODE_INKBOX_API_KEY / CLAUDE_CODE_INKBOX_SIGNING_KEY (AUT), REMOTE_INKBOX_API_KEY (driver), OPENAI_API_KEY (realtime voice), and GOOGLE_CHAT_WEBHOOK_URL. One new secret is required before the live real-model legs can run: ANTHROPIC_API_KEY — the bridged sessions are real Claude Code, so everything except the mock-channels leg needs it.

No plugin logic changes — test + workflow files only (.github/workflows/ci.yml is folded into tests.yml).